Huang Gaoshuang ; Zhou Yang ; Zhao Luying ; Gan Wenjian

Abstract

Cross-view geo-localization (CVGL), which involves matching and retrievingsatellite images to determine the geographic location of a ground image, iscrucial in GNSS-constrained scenarios. However, this task faces significantchallenges due to substantial viewpoint discrepancies, the complexity oflocalization scenarios, and the need for global localization. To address theseissues, we propose a novel CVGL framework that integrates the visionfoundational model DINOv2 with an advanced feature mixer. Our frameworkintroduces the symmetric InfoNCE loss and incorporates near-neighbor samplingand dynamic similarity sampling strategies, significantly enhancinglocalization accuracy. Experimental results show that our framework surpassesexisting methods across multiple public and self-built datasets. To furtherimprove globalscale performance, we have developed CV-Cities, a novel datasetfor global CVGL. CV-Cities includes 223,736 ground-satellite image pairs withgeolocation data, spanning sixteen cities across six continents and covering awide range of complex scenarios, providing a challenging benchmark for CVGL.The framework trained with CV-Cities demonstrates high localization accuracy invarious test cities, highlighting its strong globalization and generalizationcapabilities. Our datasets and codes are available athttps://github.com/GaoShuang98/CVCities.